compile3 0.10.1 driving_vision in mac pytest (#14911)

* compile3 0.10.1 driving_vision in mac pytest

* sync before re-executing onetime kernels

.github/workflows/benchmark.yml

@@ -45,6 +45,8 @@ jobs:
       run: |
         source /tmp/tinygrad_pytest_ci/bin/activate
         pytest -nauto --durations=20
+    - name: openpilot compile3 0.10.1 driving_vision
+      run: FLOAT16=1 CL=1 IMAGE=2 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
 
   testmacbenchmark:
     name: Mac Benchmark

tinygrad/engine/jit.py

@@ -348,6 +348,8 @@ class TinyJit(Generic[ReturnType]):
         update_depends(depends, jit_cache)
         pruned, onetime = partition(jit_cache, lambda ei: any(b in depends for b in get_out_buffers_for_ei(ei)))
         if DEBUG >= 1: print(f"pruned from {len(jit_cache)} -> {len(pruned)} kernels")
+        # sync before re-executing onetime kernels
+        for dev in set(Device[b.device] for ei in onetime for b in ei.bufs if b is not None): dev.synchronize()
         # run the onetime kernels here
         for ei in onetime:
           for b in ei.bufs: cast(Buffer, b).ensure_allocated()